Following the work of Chakrabarty et al. [2021], we performed human evaluation of the newly constructed metaphorical sentences, by comparing them against human-generated ones.

For each sentence, then, four different dimensions were evaluated, vis. **fluency**, **meaning**, **creativity**, and **metaphoricity**.

Human Evaluation Metric

This two-stage pipeline first trains a classifier to distinguish metaphorical sentences from literal ones. Sentences classified as metaphorical then enter a reconstruction stage in which metaphorical tokens are masked and recovered using T5 or BART extraction methods, or BERT or XLM-R masked language modeling.

Google

University of Michigan - Ann Arbor

Generation of new metaphors requires sentences from a corpus $$C$$ which $$MC$$ has classified as literal. A random token is masked and the masked metaphor model replaces the masked token, effectively trying to create a metaphorical text. Finally, use $$MC$$ to keep only truly transferred metaphorical texts.

Learn Before

Related

Learn After